Frustration Erupts Inside Meta Over the AI-Only Push — Staff Call It "The Gulag"

According to a report by Inc., discontent is mounting among employees at Meta (formerly Facebook) over CEO Mark Zuckerberg's aggressive AI development push. Some staff refer to being reassigned to AI projects as "The Gulag," arguing that Meta's culture—once hailed as a dream workplace—has changed drastically.

Zuckerberg has positioned AI at the core of Meta's growth and continues to pour massive resources into it, but this rapid strategic pivot has created friction within the organization. The move highlights the risk that major tech companies take on when they sacrifice organizational culture amid the AI race. That said, the reporting draws on the voices of some employees and may not necessarily reflect a company-wide consensus.

AI Has Given the US "New Power" — An Economist Analysis

The Economist argues that advances in AI technology have handed the United States geopolitical and economic advantages. It points out that a combination of control over the semiconductor supply chain, a concentration of AI talent, and the market dominance of leading AI companies has built a position in the AI domain where the US overwhelms other nations.

According to the analysis, this "new power" functions as indirect influence exerted through technological infrastructure, on top of conventional military and economic superiority. Meanwhile, countries outside the US—notably the EU and China—are strengthening their own AI strategies from a sovereignty standpoint, and it remains highly uncertain whether this advantage can be sustained in the long term.

Microsoft Discovers a New Type of Malware Targeting Cryptocurrency

Microsoft has announced the detection of a new malware strain called "Crypto Clipper," which self-propagates via USB and steals cryptocurrency credentials. The malware monitors clipboard contents, and when it detects patterns matching wallet addresses or seed phrases, it sends them to the attacker. It communicates over the Tor network and also functions as a lightweight backdoor that does not depend on traditional C2 infrastructure.

Security researchers note that this malware represents a new threat model that combines financially motivated information-stealing malware with remote code execution capabilities. While not an AI/ML story per se, it stands out as part of a broader trend in which attacks targeting cryptocurrency and digital assets are growing more sophisticated as AI technology accelerates.

Moebius — A Lightweight Framework Achieving 10B-Class Image Inpainting with 0.2B Parameters

Moebius, which has been drawing attention on Hugging Face Daily Papers, achieves image inpainting quality on par with or better than industrial-grade SOTA models with 10B+ parameters—such as FLUX.1-Fill-Dev—using only 0.22B parameters. With inference more than 15 times faster, it offers a fascinating approach that challenges "scale supremacy."

The research team claims that the synergy between architecture design and knowledge distillation breaks the "impossible triangle" of low parameter count, fast inference, and high quality. As an example pointing toward efficient AI development that does not depend on scaling laws, it is expected to spur discussion in the research community.

Limits of Agent Benchmark Evaluation Methods — A New Paper Weighs In

"Beyond Static Leaderboards," also published on Hugging Face Daily Papers, argues that there are fundamental problems with how LLM agents are evaluated. The paper consolidates 14 parallel implementation studies of MCP-based industrial agent benchmarks and demonstrates that rankings based on aggregate scores are unstable in out-of-distribution settings.

As an alternative, it proposes evaluation via "predictive validity"—the rank correlation between in-sample and out-of-sample settings—and presents a 12-layer measurement framework covering the diverse dimensions of deployment environments. The argument that a single score is insufficient for evaluating AI agents is a key point that pushes for a rethinking of benchmark design.

---